The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill architecture implements a state-sharing mechanism via the window.zcode.setWidgetState API. Data stored in the modelContent field of this state is explicitly designed to be included as untrusted input in the agent's next turn. This feedback loop creates a surface for indirect prompt injection, as malicious data or instructions introduced into the visualization's state (via user interaction or code logic) can influence subsequent agent responses. Furthermore, the sendFollowUpMessage API allows the visualization's code to programmatically trigger new agent prompts.
Ingestion points: Visualization state (modelContent) and follow-up prompts (sendFollowUpMessage) entering the agent's context (SKILL.md, api.md).
Boundary markers: The documentation notes that modelContent is treated as untrusted data, though it lacks specific instructions for the agent to use boundary markers when processing it.
Capability inventory: The skill allows file-system writes, subprocess execution (scripts/render.py), and network requests to whitelisted CDNs.
Sanitization: While the instructions direct the agent to escape literals during code generation, there is no verification or sanitization mechanism for the data returned from the interactive widget.
[DYNAMIC_EXECUTION]: The skill's primary function is to generate HTML, CSS, and JavaScript fragments at runtime. These fragments are written to the local file system and executed within a sandboxed environment, which is a high-capability pattern characteristic of dynamic code execution.
[COMMAND_EXECUTION]: The instructions require the agent to execute a local Python utility (scripts/render.py) to wrap fragments into standalone files or serve them for inspection.
[PERSISTENCE]: The skill employs a prompt-persistence technique by instructing the agent to include a "Reload the full visualize skill" command in every compaction summary, ensuring that its extensive instruction set is maintained in the active conversation context.

visualize